A multi-medium sound field manipulation system based on an AI algorithm

Abstract

This invention discloses a multi-media sound field manipulation system based on AI algorithms, belonging to the field of intelligent technology. It includes a data acquisition module, a sound field modeling module, an AI optimization control module, an adaptive adjustment module, and a hardware execution module. The data acquisition module samples and acquires acoustic signals from air, water, and solid media. The sound field modeling module combines the finite-difference time-domain method with graph neural networks to accurately simulate sound wave propagation. The AI ​​optimization control module integrates near-end strategy optimization and differential evolution algorithms to generate optimal excitation parameters. The adaptive adjustment module is based on model transfer and comparative learning mechanisms. The hardware execution module uses a piezoelectric transducer array to generate a controllable sound field. The feedback evaluation module uses laser vibration measurement and a microphone array to transmit measurement data back, achieving closed-loop correction of control parameters. Beneficial effects include high-precision sound wave modeling, intelligent control parameter optimization, and dynamic adaptive adjustment capabilities, improving the real-time performance and stability of sound field manipulation in multi-media environments.

Description

Technical Field

[0001] This invention relates to the field of intelligence, and more specifically, to a multi-media sound field control system based on AI algorithms. Background Technology

[0002] Multi-media sound field manipulation systems utilize the propagation characteristics of sound waves in different media, and are widely used in acoustics, communications, medicine, environmental monitoring, and other fields. With the rapid development of artificial intelligence (AI) algorithms, traditional sound field control methods are gradually evolving towards intelligence and automation. AI-based multi-media sound field manipulation systems leverage advanced technologies such as deep learning and machine learning to achieve precise control of the sound field by real-time monitoring, analysis, and optimization of the propagation path and effects of sound waves in multiple media. The background technology development of this system began with in-depth research into the propagation characteristics of sound waves in the acoustic field, and was gradually applied to the manipulation of sound waves in different media (such as air, water, and solid materials). Traditional sound field manipulation relies heavily on physical models and empirical design, which has limitations. The introduction of AI algorithms overcomes these limitations. Through big data analysis and model training, the sound field manipulation system can automatically adapt to complex environments and precisely adjust parameters such as sound wave frequency, intensity, and direction.

[0003] In existing technologies, many AI algorithms rely on training with large amounts of experimental data, which are often limited by the experimental environment and lack sufficient diversity. Therefore, the generalization ability of AI models is poor, and they may not be able to effectively cope with changes in the sound field in complex or unknown media environments, resulting in unsatisfactory control effects. Although AI algorithms can perform well under certain fixed conditions, in practical applications, the system's adaptability is insufficient due to changes in the media environment (such as variations in temperature, pressure, humidity, etc.). Existing technologies have not yet achieved sufficiently intelligent dynamic adjustment capabilities, and may not be able to maintain optimal sound field control effects under changing environmental conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-media sound field control system based on AI algorithms to address the problems mentioned in the background: In existing technologies, multiple AI algorithms rely on a large amount of experimental data for training, but this data is often limited by the experimental environment and lacks sufficient diversity. Therefore, the generalization ability of AI models is poor, and in complex or unknown media environments, they may not be able to effectively cope with changes in the sound field, resulting in unsatisfactory control effects. Although AI algorithms can perform well under certain fixed conditions, in practical applications, due to changes in the media environment (such as changes in temperature, pressure, humidity, etc.), the system's adaptability is insufficient. Existing technologies have not yet achieved sufficiently intelligent dynamic adjustment capabilities, and may not be able to maintain optimal sound field control effects under changing environmental conditions.

[0005] Technical solution: The AI-based multi-media sound field manipulation system includes a data acquisition module, a sound field modeling module, an AI optimization control module, an adaptive adjustment module, a hardware execution module, and a feedback evaluation module;

[0006] The data acquisition module acquires acoustic signals from air, water, and solid media through multi-channel synchronous sampling; the sound field modeling module uses a finite difference time domain method combined with a graph neural network to calculate sound wave propagation characteristics; the AI ​​optimization control module combines a near-end strategy optimization algorithm and a differential evolution algorithm to calculate sound field control parameters; the adaptive adjustment module uses a model transfer method to optimize the weight parameters of the AI ​​optimization control module; the hardware execution module excites the target sound field based on a piezoelectric transducer array; and the feedback evaluation module uses a laser Doppler vibrometer and a microphone array to acquire sound field measurement data and correct the control parameters.

[0007] Preferably, the sound field modeling module includes a physical modeling unit and a data-driven modeling unit;

[0008] The physical modeling unit calculates the sound field distribution in a non-uniform medium based on the finite difference time domain method.

[0009] The data-driven modeling unit uses a graph neural network to learn the influence of medium characteristics on sound wave propagation.

[0010] The physical modeling unit employs absorbing boundary conditions to reduce reflections at the boundary of the computational region.

[0011] The data-driven modeling unit uses a multi-layer graph convolutional network to extract media topology features.

[0012] Preferably, in the graph convolutional network of the data-driven modeling unit, the input layer uses an adjacency matrix to represent the spatial topological relationship of the multi-medium sound field; the hidden layer uses a self-attention mechanism to adjust the propagation weights between media; the output layer calculates the sound field distribution and compares it with the calculation results in the finite difference time domain, and optimizes the network parameters through gradient descent.

[0013] Preferably, the self-attention mechanism of the hidden layer adopts a multi-head attention structure;

[0014] The multi-head attention structure optimizes feature extraction by computing different weight matrices in parallel.

[0015] The weight matrix adopts a dynamic weight update strategy based on position encoding;

[0016] The dynamic weight update strategy adaptively adjusts attention allocation based on the similarity between neighboring nodes.

[0017] Preferably, the AI ​​optimization control module includes a reinforcement learning unit and an evolutionary algorithm unit;

[0018] The reinforcement learning unit uses a near-end policy optimization algorithm to train the agent to calculate the optimal sound field control strategy.

[0019] The evolutionary algorithm unit uses the differential evolution algorithm to optimize the hyperparameters of the reinforcement learning unit;

[0020] The optimization objectives of the reinforcement learning unit include sound pressure level deviation, energy utilization, and computational cost.

[0021] The evolutionary algorithm unit adjusts the search step size based on an adaptive mutation factor.

[0022] Preferably, the reinforcement learning unit employs a hierarchical reward mechanism to optimize the sound field control strategy;

[0023] The main reward is calculated based on the sound pressure level deviation in the target area to determine the gradient optimization direction;

[0024] The auxiliary reward is adjusted based on the sound field energy utilization rate and computational complexity.

[0025] The reward function employs a dynamic weight allocation strategy;

[0026] The dynamic weight allocation strategy is based on KL divergence calculation to determine the magnitude of change between the current strategy and historical strategies.

[0027] Preferably, the reinforcement learning unit employs self-supervised learning to enhance generalization ability;

[0028] The self-supervised learning method uses a contrastive learning approach to construct positive and negative sample pairs.

[0029] The positive sample pairs are composed of optimal sound field control parameters under similar environments;

[0030] The negative sample pairs are composed of mistuned parameters under different media conditions;

[0031] By maximizing positive sample similarity and minimizing negative sample similarity.

[0032] Preferably, the evolutionary algorithm unit uses a hierarchical differential evolution strategy to optimize the hyperparameters of the reinforcement learning unit;

[0033] The hierarchical differential evolution strategy includes individual mutation, population selection, and hierarchical crossover.

[0034] The individual variation is adaptively adjusted based on the gradient information of the optimal individual;

[0035] The population selection employs an elite retention strategy to inherit the optimal hyperparameters.

[0036] The hierarchical cross-referencing is based on a multi-objective optimization strategy to search for the Pareto optimal solution.

[0037] Preferably, the hierarchical differential evolution strategy is combined with the neural architecture search method to optimize the hyperparameter search space;

[0038] The neural architecture search method employs a gradient-based search strategy to optimize the network structure.

[0039] The search strategy optimizes the architecture through second-order gradient updates.

[0040] The neural network structure of the reinforcement learning unit is selected from fully connected networks, convolutional networks, and transformer structures based on a search strategy.

[0041] Compared with the prior art, the advantages of this invention are:

[0042] (1) This invention combines the finite difference time domain method (FDTD) and graph neural network (GNN) to optimize the accuracy of sound field modeling in multi-media environments and overcome the problem of insufficient simulation accuracy of traditional methods in complex media.

[0043] (2) This invention improves the adaptability of the control policy through proximal policy optimization (PPO) and differential evolution algorithm (DE), and enhances the generalization ability in different media environments by adopting self-supervised learning, thereby improving training efficiency and accuracy.

[0044] (3) By introducing multi-sensor feedback of laser Doppler vibration meter and microphone array, combined with error correction, the present invention improves the closed-loop stability and accuracy of sound field control and enhances the robustness of the system in actual environment. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of a multi-media sound field control system based on AI algorithms according to the present invention. Detailed Implementation

[0046] For examples, please refer to Figure 1 A multi-media sound field manipulation system based on AI algorithms includes a data acquisition module, a sound field modeling module, an AI optimization control module, an adaptive adjustment module, a hardware execution module, and a feedback evaluation module.

[0047] The system comprises the following modules: a data acquisition module that acquires acoustic signals from air, water, and solid media through multi-channel synchronous sampling; a sound field modeling module that uses a finite difference time domain method combined with a graph neural network to calculate sound wave propagation characteristics; an AI optimization control module that combines a near-end strategy optimization algorithm with a differential evolution algorithm to calculate sound field control parameters; an adaptive adjustment module that uses a model transfer method to optimize the weight parameters of the AI ​​optimization control module; a hardware execution module that excites the target sound field based on a piezoelectric transducer array; and a feedback evaluation module that uses a laser Doppler vibrometer and a microphone array to acquire sound field measurement data and correct control parameters.

[0048] Specifically, the data acquisition module is based on multi-channel synchronous sampling technology to collect acoustic signals in air, water, and solid media. It uses a high-precision analog-to-digital converter to digitize the minute sound pressure signals and combines environmental sensing sensors (temperature, humidity, and density sensors) to synchronously acquire the physical properties of the media, providing basic data for modeling.

[0049] The sound field modeling module includes a physical modeling unit and a data-driven modeling unit. The former calculates the sound wave propagation path and reflection and refraction effects in a non-uniform medium based on the finite difference time domain (FDTD) method, and uses the absorbing boundary condition (PML) to weaken boundary reflections. The latter learns the topological properties of the medium through a graph neural network (GNN). The input layer expresses spatial connectivity relationships with an adjacency matrix, the hidden layer introduces a multi-head self-attention mechanism to adjust the propagation weights between media, and the output layer compares the modeling results with the FDTD results and jointly optimizes the weight parameters through a gradient descent algorithm.

[0050] The AI ​​optimization control module includes a reinforcement learning unit and an evolutionary algorithm unit. The reinforcement learning unit uses the Proximal Policy Optimization (PPO) algorithm to train the control agent. Its state space is the current sound field state, and its action space is the combination of excitation parameters (such as transducer frequency, phase, amplitude, etc.). The reward function combines the main reward (sound pressure deviation in the target area) and the auxiliary reward (energy utilization rate and computational cost) and uses a dynamic weight allocation strategy controlled by KL divergence. The evolutionary algorithm unit uses the Differential Evolution (DE) algorithm to optimize the reinforcement learning hyperparameters and introduces a hierarchical differential evolution strategy, including individual mutation (adaptively adjusting the mutation amplitude based on gradient information), population selection (using an elite retention strategy to inherit excellent parameters), and hierarchical crossover (multi-objective optimization to obtain the Pareto optimal solution).

[0051] The adaptive adjustment module optimizes the weights of the AI ​​module through a model transfer mechanism. When the acoustic medium, target area, or transducer array configuration changes, the system uses historical model transfer learning to quickly converge to the new optimal control strategy. It also combines self-supervised contrastive learning to enhance generalization ability. Positive sample pairs come from the optimal control parameters of similar medium environments, while negative sample pairs come from mistuned parameters. Transfer learning is achieved by maximizing the similarity of positive samples and minimizing the similarity of negative samples.

[0052] The hardware execution module uses a piezoelectric transducer array to generate adjustable sound waves. The array layout is adaptively optimized in terms of density and direction based on the modeling results. The drive signal is modulated in real time by a high-precision waveform controller and fine-tuned based on real-time feedback to ensure that the generated sound field covers the target area.

[0053] The feedback evaluation module acquires sound field information in real time through a laser Doppler vibrometer and a three-dimensional microphone array, constructs an error evaluation function, and sends a correction signal back to the AI ​​control module to achieve closed-loop control. The measurement results are used to evaluate parameters such as sound pressure level, phase shift, and energy distribution, thereby further improving the control accuracy.

[0054] The sound field modeling module includes a physical modeling unit and a data-driven modeling unit;

[0055] The physical modeling unit calculates the sound field distribution in a non-uniform medium based on the finite difference time domain method.

[0056] The data-driven modeling unit uses a graph neural network to learn the influence of medium properties on sound wave propagation;

[0057] The physical modeling unit uses absorbing boundary conditions to reduce reflections at the boundary of the computational domain;

[0058] The data-driven modeling unit uses a multi-layer graph convolutional network to extract media topology features.

[0059] Specifically, the sound field modeling module adopts the following steps:

[0060] A1. Use the FDTD method to generate a mesh model and set medium parameters (density, sound velocity, etc.) and boundary absorption conditions;

[0061] A2. Input the initial excitation source parameters and iteratively calculate the sound field changes at each time step;

[0062] A3. Construct a topology graph model and establish the connection relationships between media nodes to form an adjacency matrix;

[0063] A4. Graph neural networks extract the topological features of the medium and calculate the sound propagation path adjustment coefficient through multi-layer graph convolution (GCN) and attention mechanism;

[0064] A5. Cross-validate with the FDTD calculation results, and use gradient descent to optimize model parameters and improve modeling accuracy.

[0065] In the graph convolutional network of the data-driven modeling unit, the input layer uses an adjacency matrix to represent the spatial topological relationship of the multi-medium sound field; the hidden layer uses a self-attention mechanism to adjust the propagation weights between media; the output layer calculates the sound field distribution and compares it with the results calculated in the finite difference time domain, and optimizes the network parameters through gradient descent.

[0066] Specifically, the self-attention mechanism of graph neural networks uses the following formula to adjust weights:

[0067]

[0068] in: For node feature vectors, To query the weight matrix of the key, For dimension normalization factor, This is a position encoding function.

[0069] The self-attention mechanism of the hidden layer adopts a multi-head attention structure;

[0070] Multi-head attention structures optimize feature extraction by computing different weight matrices in parallel;

[0071] The weight matrix employs a position-encoding-based dynamic weight update strategy;

[0072] The dynamic weight update strategy adaptively adjusts attention allocation based on the similarity between neighboring nodes.

[0073] Specifically, the reward function for a reinforcement learning unit adopts the following structure:

[0074]

[0075] in: For target sound pressure, This is the actual sound pressure level. For energy efficiency, To calculate resource consumption, As a weighting factor, it is dynamically adjusted; To prevent tiny constants with a denominator of zero.

[0076] The AI ​​optimization control module includes a reinforcement learning unit and an evolutionary algorithm unit;

[0077] The reinforcement learning unit uses a proximal policy optimization algorithm to train the agent to calculate the optimal sound field control strategy;

[0078] The evolutionary algorithm unit uses the differential evolution algorithm to optimize the hyperparameters of the reinforcement learning unit;

[0079] The optimization objectives of reinforcement learning units include sound pressure level deviation, energy utilization, and computational cost;

[0080] The evolutionary algorithm unit adjusts the search step size based on an adaptive mutation factor.

[0081] Specifically, the hierarchical evolution process in the differential evolution algorithm is as follows:

[0082] B1. Initialize the population and set the individual parameter vectors;

[0083] B2. Adaptively adjust the coefficient of variation and crossover probability according to the gradient change of the best individual in each iteration;

[0084] B3. Use the Pareto optimal solution to score and filter candidate parameters;

[0085] B4. Introduce neural architecture search and combine candidate architecture space (fully connected, convolution, transformer) to optimize the AI ​​control network structure and improve policy convergence speed.

[0086] The reinforcement learning unit employs a tiered reward mechanism to optimize the sound field control strategy;

[0087] The main reward is calculated based on the gradient optimization direction of the sound pressure level deviation in the target area;

[0088] The auxiliary reward is adjusted based on the sound field energy utilization rate and computational complexity.

[0089] The reward function employs a dynamic weight allocation strategy;

[0090] The dynamic weight allocation strategy is based on KL divergence to calculate the magnitude of change between the current strategy and historical strategies.

[0091] Specifically, it supports multimodal input and remote linkage operation. After inputting the task target through the remote control platform, the system can adaptively configure the sound field control scheme under different media structures, and supports real-time three-dimensional sound map rendering and multi-task area sound focus control.

[0092] The reinforcement learning unit employs self-supervised learning to enhance generalization ability;

[0093] Among them, self-supervised learning uses a contrastive learning method to construct positive and negative sample pairs;

[0094] Positive sample pairs are composed of optimal sound field control parameters under similar environments;

[0095] Negative sample pairs are composed of mistuned parameters under different media conditions;

[0096] By maximizing positive sample similarity and minimizing negative sample similarity.

[0097] Specifically, a joint loss function formula for sound field modeling and optimization control in this invention has the following form:

[0098]

[0099] in: These are the sound pressure levels predicted by the physical model and the GNN, respectively.

[0100] KL represents the KL divergence of the control policy change. The gradient of the reward function with respect to the network parameters.

[0101] This is the weighting factor for the loss term.

[0102] The evolutionary algorithm unit uses a hierarchical differential evolution strategy to optimize the hyperparameters of the reinforcement learning unit;

[0103] Hierarchical differential evolution strategies include individual variation, population selection, and hierarchical crossover;

[0104] Individual variation adaptively adjusts the variation magnitude based on the gradient information of the optimal individual;

[0105] Population selection employs an elite retention strategy to inherit the optimal hyperparameters;

[0106] Hierarchical cross-referencing is based on a multi-objective optimization strategy to search for the Pareto optimal solution.

[0107] Specifically, a formula for a multi-media dynamic control sound focusing index:

[0108]

[0109] in, Let be the sound pressure at the i-th sampling point; This is the maximum sound pressure level. The standard deviation of sound pressure. This is the adjustment coefficient for regulating uniformity.

[0110] The hierarchical differential evolution strategy combined with the neural architecture search method optimizes the hyperparameter search space;

[0111] Among them, the neural architecture search method uses a gradient-based search strategy to optimize the network structure;

[0112] The search strategy optimizes the architecture through second-order gradient updates.

[0113] The neural network structure of the reinforcement learning unit is selected from fully connected networks, convolutional networks, and transformer structures based on a search strategy.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI algorithm-based multi-medium sound field manipulation system, characterized in that, The AI-based multi-media sound field manipulation system includes a data acquisition module, a sound field modeling module, an AI optimization control module, an adaptive adjustment module, a hardware execution module, and a feedback evaluation module. The data acquisition module acquires acoustic signals from air, water, and solid media through multi-channel synchronous sampling; the sound field modeling module uses a finite difference time-domain method combined with a graph neural network to calculate sound wave propagation characteristics; the AI ​​optimization control module combines a near-end strategy optimization algorithm and a differential evolution algorithm to calculate sound field control parameters; the adaptive adjustment module uses a model transfer method to optimize the weight parameters of the AI ​​optimization control module; the hardware execution module excites the target sound field based on a piezoelectric transducer array; and the feedback evaluation module uses a laser Doppler vibrometer and a microphone array to acquire sound field measurement data and correct the control parameters. The sound field modeling module includes a physical modeling unit and a data-driven modeling unit; The physical modeling unit calculates the sound field distribution in a non-uniform medium based on the finite difference time domain method. The data-driven modeling unit uses a graph neural network to learn the influence of medium characteristics on sound wave propagation. The physical modeling unit employs absorbing boundary conditions to reduce reflections at the boundary of the computational region. The data-driven modeling unit uses a multi-layer graph convolutional network to extract media topology features; In the graph convolutional network of the data-driven modeling unit, the input layer uses an adjacency matrix to represent the spatial topological relationship of the multi-medium sound field; the hidden layer uses a self-attention mechanism to adjust the propagation weights between media; the output layer calculates the sound field distribution and compares it with the calculation results in the finite difference time domain, and optimizes the network parameters through gradient descent. The self-attention mechanism of the hidden layer adopts a multi-head attention structure; The multi-head attention structure optimizes feature extraction by computing different weight matrices in parallel. The weight matrix adopts a dynamic weight update strategy based on position encoding; The dynamic weight update strategy adaptively adjusts attention allocation based on the similarity between neighboring nodes.

2. The multi-medium sound field manipulation system of claim 1, wherein, The AI ​​optimization control module includes a reinforcement learning unit and an evolutionary algorithm unit; The reinforcement learning unit uses a near-end policy optimization algorithm to train the agent to calculate the optimal sound field control strategy. The evolutionary algorithm unit uses the differential evolution algorithm to optimize the hyperparameters of the reinforcement learning unit; The optimization objectives of the reinforcement learning unit include sound pressure level deviation, energy utilization, and computational cost. The evolutionary algorithm unit adjusts the search step size based on an adaptive mutation factor.

3. The multi-media sound field manipulation system according to claim 2, characterized in that, The reinforcement learning unit employs a hierarchical reward mechanism to optimize the sound field control strategy using a reward function. The hierarchical reward includes a primary reward and an auxiliary reward. The main reward is calculated based on the sound pressure level deviation in the target area to determine the gradient optimization direction; The auxiliary reward is adjusted based on the sound field energy utilization rate and computational complexity. The reward function employs a dynamic weight allocation strategy; The dynamic weight allocation strategy is based on KL divergence calculation to determine the magnitude of change between the current strategy and historical strategies.

4. The multi-medium sound field manipulation system of claim 3, wherein, The reinforcement learning unit employs self-supervised learning to enhance generalization ability; The self-supervised learning method uses a contrastive learning approach to construct positive and negative sample pairs. The positive sample pairs are composed of optimal sound field control parameters under similar environments; The negative sample pairs are composed of mistuned parameters under different media conditions; By maximizing positive sample similarity and minimizing negative sample similarity.

5. The multi-medium sound field manipulation system of claim 2, wherein, The evolutionary algorithm unit uses a hierarchical differential evolution strategy to optimize the hyperparameters of the reinforcement learning unit; The hierarchical differential evolution strategy includes individual mutation, population selection, and hierarchical crossover. The individual variation is adaptively adjusted based on the gradient information of the optimal individual; The population selection employs an elite retention strategy to inherit the optimal hyperparameters. The hierarchical cross-referencing is based on a multi-objective optimization strategy to search for the Pareto optimal solution.

6. The multi-medium sound field manipulation system of claim 5, wherein, The hierarchical differential evolution strategy, combined with the neural architecture search method, optimizes the hyperparameter search space. The neural architecture search method employs a gradient-based search strategy to optimize the network structure. The search strategy optimizes the architecture through second-order gradient updates. The neural network structure of the reinforcement learning unit is selected from fully connected networks, convolutional networks, and transformer structures based on a search strategy.

Images